Abstract
The aim of this project is to articulate a theory of first language acquisition, and to demonstrate this theory by means of a computational model that simulates the acquisition of early Dutch syntax. First language acquisition is a controversial issue in theoretical linguistics. There are two opposing points of view. The generative tradition emphasizes the abstractness of grammar and concludes that language can only be learned on the basis of innate linguistic knowledge. An opposing view, the so-called usage-based approach, holds that child language is less abstract and well learnable. The eventual abstractness of adult grammar is doubted or put aside. We will argue that this discussion is based on a false dilemma. Longitudinal studies of child language development in simple descriptive terms show that the adult grammar is the final outcome of a stepwise development of abstractness. It involves a hierarchy of intermediate grammars in a predictable order. The proposed project will demonstrate this with algorithmic precision without invoking innate linguistic knowledge. At each stage in grammatical development, the child effectively simplifies its input by applying a data-reduction operation appropriate for that state of its grammar. In the proposed project we will describe the steps involved in early Dutch syntax acquisition in some detail and demonstrate how the child?s simplifications of the adult system can be simulated and how they explain the acquisition steps. Data-reduction strategies and machine-learning principles will implement these steps in a computational model and test them on data from the CHILDES corpus. A crucial test for the algorithm will be the acquisition of grammatical markings that have a double dependency: a sentential scope function (negation, question and illocution function) as well as a phrasal internal function (marking a predicate or an argument).